Receptive Field-Based All-Optical Spiking Neural Network for Image Processing

Abstract

We report on a novel structure of a receptive field (RF)-based multi-layer all-optical neural network using a micropillar laser with a saturable absorber (SA) for image processing. From the perspective of biological vision, the realization of image processing based on the RF provides the biological rationality for the machine vision implemented by the spiking neural network (SNN). By exploiting the fast physical mechanisms of gain and absorption in the SA laser, the photonic spike-timing-dependent plasticity (STDP) curves are achieved to train the weights. Here, the source image pixels are mapped into the temporal information of spike trains injected into the neural network through the temporal coding method called time-to-first-spike encoding. Different source images are processed and tested by the proposed photonic SNN. Simulation results show that our proposed system can process not only simple binary images but also complex color images under the adjustment of STDP rules. When considering the robustness, we demonstrate the tolerance of the image segmentation to the time jitter. These results indicate that our proposed photonic SNN can achieve high-resolution processing of complex source images. Additionally, the time-multiplexing technique can be further adopted to simplify the RF structure, which is expected to reduce the complexity of the whole system, thus facilitating physical applications. Our work offers the prospect for a high-speed photonic spiking platform for image processing.